Question

A retail company trains a fraud-detection model in BigQuery ML using engineered features (rolling 7-day transaction counts, average purchase amount) computed from historical data. The data science team now needs to serve this model for real-time predictions at checkout, where the same features must be looked up with low latency. They also want to guarantee that the feature values used during training match those used at serving time to avoid training-serving skew. Which approach best meets these requirements?

Accepted Answer

Store the engineered features in Vertex AI Feature Store, using it as the offline store for training data and the online store for low-latency serving lookups.

Answer

Recompute all features on the fly at prediction time within the application layer using direct BigQuery queries against raw transaction tables.

Answer

Export the training features to a Cloud Storage CSV file and load them into a Memorystore Redis cache accessed by the application at serving time.

Answer

Keep the features only in BigQuery and enable BI Engine to accelerate the point lookups needed during real-time checkout predictions.